New DevHeads get a 320-point leaderboard boost when joining the DevHeads IoT Integration Community. In addition to learning and advising, active community leaders are rewarded with community recognition and free tech stuff. Start your Legendary Collaboration now!
I’m building a CNN model to run on an STM32F746NG Discovery board following the “TensorFlow Lite for Microcontrollers” tutorials and the TinyML book. I know that the supported TensorFlow-Keras functions are listed in the all_ops_resolver.cc file (https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/micro/all_ops_resolver.cc), but I’m having trouble with the Flatten() function. It doesn’t seem to be listed, even though it’s such a basic operation. I’m using other functions from the list, and my model looks like this:
model = models.Sequential()
model.add(layers.Conv2D(16, (3, 3), activation=’relu’, input_shape=(36, 36, 1)))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(32, (3, 3), activation=’relu’, input_shape=(36, 36, 1)))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(64, (3, 3), activation=’relu’, input_shape=(36, 36, 1)))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Flatten()) # Potential issue here
model.add(layers.Dense(8, activation=’softmax’))
model.add(layers.Dense(2))
model.summary()
model.compile(optimizer=’adam’,
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=[‘accuracy’])
I get a segmentation fault when running the model on the STM32F746NG board, no matter how much memory I allocate. I suspect the issue might be related to the Flatten() function, but I wanted to ask: is Flatten() supported by TensorFlow Lite for Microcontrollers? If not, does it go by a different name or is there an alternative I should use? The segmentation fault seems to be triggered in the process, and I’m trying to figure out if it’s specific to this function or something else in my setup.
Has anyone encountered this issue or found a workaround for Flatten() on the STM32F746NG with TensorFlow Lite?
Since `Flatten()` is a simple operation that reshapes the data from a multi dimensional array to a single dimensional array, you can replace it with a Reshape operation, which is supported in TensorFlow Lite for Microcontrollers
You can reshape the data without relying on `Flatten()` by adding a Reshape layer.
This might be helpful
0reshaping.txt
@wafa_ath You’re encountering a segmentation fault when running your `CNN` model on the `STM32F746NG` board using `TensorFlow` Lite, specifically when using the `Flatten()` function. Based on your description, the issue might `stem` from the fact that `Flatten()` is not explicitly supported in `TensorFlow Lite` for Microcontrollers, as it’s not included in the `all_ops_resolver.cc` file. However `Flatten()` is a common operation in standard `TensorFlow` models, it’s not supported by `TensorFlow Lite` for Microcontrollers. However, you can achieve the same `functionality` by replacing the `Flatten()` layer with a `Reshape layer`, which is supported in `TensorFlow Lite` for Microcontrollers and performs the equivalent `operation—flattening` a `multi-dimensional` tensor into a `1D array`.
Here’s how you can modify your model to replace the `Flatten()` layer:
“`
model = models.Sequential()
model.add(layers.Conv2D(16, (3, 3), activation=’relu’, input_shape=(36, 36, 1)))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(32, (3, 3), activation=’relu’))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(64, (3, 3), activation=’relu’))
model.add(layers.MaxPooling2D((2, 2)))
# Replace Flatten() with Reshape
model.add(layers.Reshape((-1,))) # This layer automatically flattens the tensor
model.add(layers.Dense(8, activation=’softmax’))
model.add(layers.Dense(2))
model.summary()
model.compile(optimizer=’adam’,
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=[‘accuracy’])
“`
Thanks so much for the suggestion! I tried replacing Flatten() with Reshape((-1,)), but I’m still getting the same segmentation fault after that layer on the STM32F746NG. I also checked and adjusted memory allocation, but no luck so far.
I’m wondering if the issue might be related to the overall model size or how the layers are handled during deployment. If anyone has other ideas or suggestions on how to fix this, I’d love to hear them!
Thanks again for the help—I really appreciate it!
CONTRIBUTE TO THIS THREAD